☀️ TRENDING AI NEWS
🤖 MiniMax open-sources M2.7 - a self-evolving agent that scores 56% on SWE-Pro coding benchmarks
🏢 AI companies pour money into policy papers and thinktanks to fix their public image problem
🎵 Real musicians are finding AI-generated imposters on Spotify, fraudulently streaming under their names
⚠️ Anthropic faces scepticism over whether its decision to withhold Mythos was responsibility or marketing
Something quietly shifted in the open-source AI landscape this weekend - and if you're building with code agents, you'll want to pay attention.
MiniMax just dropped the weights for M2.7 - a model that doesn't just write code, it helps build itself. Meanwhile, AI's biggest labs are quietly spending to fix a perception problem the public didn't ask them about. And real musicians are waking up to find AI clones of themselves streaming on Spotify. Let's get into it.
🤓 AI Trivia
What does "SWE-bench" actually test AI models on?
💻 Generating creative writing under time pressure
💻 Solving real GitHub software engineering issues
💻 Translating code between programming languages
💻 Optimising database query performance
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
🤖 MiniMax M2.7: The Open-Source Agent That Helped Build Itself
MiniMax just open-sourced M2.7 on Hugging Face - and the headline feature is genuinely unusual. This is MiniMax's first model to actively participate in its own development cycle, meaning it was used to generate and evaluate training data for itself. That's a meaningful step beyond the standard "trained on internet data" pipeline.
56% on SWE-Pro, 57% on Terminal Bench 2
The benchmarks are solid for an open-weight model. M2.7 scores 56.22% on SWE-Pro and 57.0% on Terminal Bench 2 - two of the harder coding evaluations available right now. SWE-Pro is specifically designed to be harder than standard SWE-bench, testing multi-step reasoning on real engineering tasks rather than simple completions.
Originally announced back on March 18, 2026, the weights are now fully public. For developers evaluating open-source coding agents, this one belongs on your radar - especially if you're keeping an eye on how self-improvement loops change what's possible at the frontier.
🏢 AI Labs Are Funding Thinktanks to Win Back the Public
Here's a telling data point: polls are showing growing public disapproval of AI, and the industry's response is to fund policy papers and thinktanks. OpenAI made a surprise announcement recently - not a new model, but a policy paper calling for "a slate of people-first ideas" and a reimagining of the social contract. It reads less like a research paper and more like a PR campaign.
Spending on Narrative, Not Just Infrastructure
The Guardian's reporting reveals this is part of an aggressive, coordinated effort across major AI players to reshape the narrative. The timing is notable - public trust is eroding at exactly the moment these companies are trying to land massive government contracts and avoid regulation.
The uncomfortable question is whether funding AI policy reform research actually leads to better AI policy, or whether it mainly produces research that supports the funder's preferred conclusions. The industry's track record on self-regulation doesn't inspire confidence - but the alternative, waiting for governments to catch up, has its own risks.
⚠️ Anthropic's Mythos Bet: Responsibility or Calculated Hype?
This one connects to what we covered Thursday about Anthropic's Mythos cybersecurity model - but there's a sharp new angle worth covering. Anthropic announced it created an AI so powerful it chose not to release it to the public out of "overwhelming responsibility". The US Treasury Secretary reportedly summoned major bank heads to discuss the implications.
When Withholding Is the Marketing
Sceptics - including security researchers and AI policy analysts - are pushing back hard. The argument: announcing that you have a dangerously powerful model, then not releasing it, is the kind of move that generates enormous press coverage and investor interest without any of the scrutiny that comes with an actual public release.
It's genuinely hard to evaluate. If the model really does pose serious cybersecurity risks, not releasing it is the right call. But "trust us, it's too dangerous" is also the perfect unfalsifiable claim for a company that wants to be taken seriously as a frontier lab. The Guardian's piece captures this tension well.
🎵 AI Is Impersonating Musicians on Spotify - and Cashing In
Jazz composer and pianist Jason Moran got a call from a friend recently: there was a new record on Spotify with his name on it. "It has your name on it," bassist Burniss Earl Travis told him. "But I don't think it's you." It wasn't.
Generative AI Supercharges an Old Fraud
Fake streaming has been a problem for years - bots inflating play counts to skim royalties. But generative AI has changed the scale entirely. It's now trivially easy to generate music that sounds like a specific artist, attach their name to it, and upload it. The fraud extracts royalties that should go to real artists, while also muddying search results and discovery for listeners.
Experts the Guardian spoke to say platforms like Spotify are struggling to keep up. Detection is hard when the music is genuinely new - not a copy, just a very convincing imitation. This sits at an uncomfortable intersection of AI music fraud, music industry rights, and platform accountability - and right now, none of those are winning.
🛠️ Gen Z Is Over the AI Hype - But Still Can't Quit It
We covered Gen Z's complicated AI relationship before, but a new Gallup report adds real numbers to the story. Based on nearly 1,600 respondents ages 14 to 29 across the US, the data shows enthusiasm is falling and resentment is growing - even as many young people feel they have no choice but to keep using AI tools for school and work.
The Obligation Trap
The Verge describes it as a love-hate relationship with a very specific dynamic: disillusioned with AI, but compelled to use it anyway. The hype is wearing off for the digital-native generation that was supposed to embrace this technology most enthusiastically. What's replacing hype is something more like resigned pragmatism.
For anyone thinking about AI adoption curves and long-term user behaviour, this is worth sitting with. The generation that grew up with smartphones isn't automatically enthusiastic about AI - they're evaluating it the same way everyone else is, and finding plenty to be frustrated about.
Speaking of tools that actually deliver - if you've been thinking about building a website without the usual headache, 60sec.site uses AI to get you from idea to published site in under a minute. Worth a look if you're a developer, founder, or anyone who'd rather spend time on the work itself. And for more daily AI coverage, dailyinference.com has everything in one place.
🌎 Trivia Reveal
The answer is B - solving real GitHub software engineering issues! SWE-bench (and its harder variant SWE-Pro) presents AI models with actual open GitHub issues and asks them to produce working code patches. It's one of the toughest practical benchmarks available because there's no shortcut - the code either fixes the bug or it doesn't.
💬 Quick Question
The Gallup data on Gen Z got me thinking - where are you at with AI right now? Still enthusiastic, quietly frustrated, or somewhere in between? Hit reply and tell me - I read every response and I'm genuinely curious whether the sentiment in that report matches what people outside the survey are feeling.
That's it for today - see you tomorrow with more. If someone forwarded this to you, you can get it directly at dailyinference.com.